Akther Sayma, Saleheen Nazir, Samiei Shahin Alan, Shetty Vivek, Ertin Emre, Kumar Santosh
University of Memphis.
University of California, Los Angeles.
Proc ACM Interact Mob Wearable Ubiquitous Technol. 2019 Mar;3(1). doi: 10.1145/3314388. Epub 2019 Mar 29.
We address the open problem of reliably detecting oral health behaviors passively from wrist-worn inertial sensors. We present our model named (pronounced ) for detecting brushing and flossing behaviors, without the use of instrumented toothbrushes so that the model is applicable to brushing with still prevalent manual toothbrushes. We show that for detecting rare daily events such as toothbrushing, adopting a model that is based on identifying candidate windows based on events, rather than fixed-length timeblocks, leads to significantly higher performance. Trained and tested on 2,797 hours of sensor data collected over 192 days on 25 participants (using video annotations for ground truth labels), our brushing model achieves 100% median recall with a false positive rate of one event in every nine days of sensor wearing. The average error in estimating the start/end times of the detected event is 4.1% of the interval of the actual toothbrushing event.
我们解决了一个开放性问题，即如何通过佩戴在手腕上的惯性传感器被动可靠地检测口腔健康行为。我们提出了名为（发音为 ）的模型，用于检测刷牙和使用牙线的行为，该模型不使用装有仪器的牙刷，从而适用于仍广泛使用的手动牙刷刷牙情况。我们表明，对于检测诸如刷牙等罕见的日常事件，采用基于根据事件识别候选窗口而非固定长度时间块的模型，会带来显著更高的性能。在25名参与者192天内收集的2797小时传感器数据上进行训练和测试（使用视频注释作为地面真值标签），我们的刷牙模型实现了100%的中位数召回率，误报率为每九天一次事件。检测到的事件的开始/结束时间估计的平均误差为实际刷牙事件间隔的4.1%。
